Real-time CNN-based Segmentation Architecture for Ball Detection in a Single View Setup

Abstract

This paper considers the task of detecting the ball from a single viewpoint in the challenging but common case where the ball interacts frequently with players while being poorly contrasted with respect to the background. We propose a novel approach by formulating the problem as a segmentation task solved by an efficient CNN architecture. To take advantage of the ball dynamics, the network is fed with a pair of consecutive images. Our inference model can run in real time without the delay induced by a temporal analysis. We also show that test-time data augmentation allows for a significant increase the detection accuracy. As an additional contribution, we publicly release the dataset on which this work is based. CCS CONCEPTS • Computing methodologies → Interest point and salient region detections; Object detection; Scene understanding.

Ball 3D Localization From A Single Calibrated Image

“…This study reveals that our CNN gets the ability to correctly estimate the ball size on images where the ball is challenging to detect. Moreover, the similarity between the DeepSport testset (subject to occlusions, see samples in [29]) and on our evaluation set (free from occlusions), indirectly reveals that our method is robust to occlusions.…”

“…Ball detector BallSeg was trained as recommended in [29], with a random scaling strategy where ball size was kept between 14 and 37 pixels, corresponding to the range of ball sizes in the DeepSport dataset (see Table 1). To analyze the impact of working on smaller balls, we also considered halving the size range in specific trainings (see Section 4.5).…”

“…We also computed the relative distance error (MAE [%] ) that represent the ratio between the projection error and the distance to the cameras. To assess the ball detection performance, we used the Area under the Curve (AuC) associated to the Receiver Operating Characteristic (ROC) curve, plotting the True Positive (TP) rate against the False Positive (FP) rate for a varying detection threshold, as defined in [29]. Here, the TP rate measures the fraction of images in which the ball is correctly detected while the FP rate measures the mean number of false candidates that are detected per image.…”

“…To detect balls in the image space, we adopt the Stateof-the-Art solution BallSeg [29] that uses a segmentation approach. Specifically, the model, based on an ICNet architecture [32], is trained to predict a mask of the ball.…”

3D Ball Localization From A Single Calibrated Image

“…To detect balls in the image space, we adopt the Stateof-the-Art solution BallSeg [30] that uses a segmentation approach. Specifically, the model, based on an ICNet architecture [32], is trained to predict a mask of the ball.…”